Performance on 64 bit Linux vs 32 Bit

[Mark Nienberg]

ram wrote:
> I have been using MailScanner on 32 bit centos for quiet some time now
> on the ~25 Antispam servers which we have ( MailScanner + Postfix +
> Spamassassin + Custom spam engine ) 
> 
> Now I was trying to evaluate  64 bit Linux. Would Mailscanner perform
> any better on 64 bit linux. I personally have no first-hand experience
> of 64 bit linux, I thought of doing some research before I upgrade 

Here is a response from the archives:

Matt Kettler wrote:
 > Mark Nienberg wrote:
 >> I've seen comments on this list that the x86_64 didn't seem to make much
 >> difference and I admit it is simpler to use the plain x86 version, but
 >> it bothers me a little to intentionally not use the software that is
 >> specifically configured for the chip.
 >
 > Why does it bother you?
 >
 > Theoretically x86-64 should be slightly slower for most uses unless you:
 >
 > 1) have a process that needs > 4gb of virtual address space
 >     -or-
 > 2) does a lot of 64 bit math that can't be performed with SSE
 >
 > The ability to have huge processes and large amounts of physical ram is the
 > primary benefit of using a 64 bit computing architecture. The drawback is that
 > pointers become larger, taking up more memory, and causing more memory I/O than
 > would be needed if the app was 32bit. Unless you're actually using the larger
 > memory space you're increasing overhead without any benefit whatsoever. Very few
 > apps have such large memory footprints outside the realm of scientific
 > simulation or massive database crunching.
 >
 >
 > The other benefit of a 64bit computing architecture is the ability to do 64 bit
 > math operations in one instruction instead of a series of 32 bit operations.
 > However, very few applications regularly have any use for 64 bit operations
 > outside of crypto, some games, and high-end engineering/physics. Even these
 > regularly get their needs filled by using SSE, so the 64-bit math benefit is
 > very limited.
 >
 > There's some benefit here to apps using 64-bit file offsets or 64 bit time
 > format, but I've never seen a "regular" application where either kind of
 > calculation was performed often enough to have a noticeable impact on
 > performance. Some scientific simulations may do a lot of 64bit time
 > calculations, but most of those could readily use SSE for it.